Article, Vision Chip

From: Lloyd G. Rasmussen (lras@loc.gov)
Date: Thu Sep 11 1997 - 09:09:42 PDT


Don't anybody get too excited. But this shows how understanding of
biology is working its way into electronics, and at some point this
will become useful to us.

This is another article from one of my favorite places, Electronic
Engineering Times, copyright CMP Publications.

  
              Vision chip's circuitry modeled on eye and brai
                                     
                            By R. Colin Johnso
                                     
   BOUYGES, France -- Mimicking the human eye's neural networks and th
   brain, an electronic "eye" on a chip can now handle tasks that rang
   from reading sign language to avoiding collisions
  
   The product of 10 years of research, the Generic Visual Perceptio
   Processor (GVPP) was developed here by the Bureau Etudes Vision (BEV
   Stockplus, said Patrick Pirim, the chip's inventor and the bureau'
   principal scientist
  
   "Now," said BEV Stockplus president Igor Marie De l'Isle, "we want t
   bring this extraordinary chip to the U.S. market, because the detai
   with which visual streams can be recognized enables new application
   not possible before.
  
   For instance, he said, sign language involves very detailed han
   gestures that have so far been impervious to traditional recognitio
   techniques. The company also claims to benefit traditiona
   pattern-recognition applications, such as military target acquisitio
   and fire control, as well as automobile-collision avoidance, adaptiv
   cruise control and automatic "asleep-at-the-wheel" alarms
  
   Input to the electronic eye can come from video, infrared or rada
   signals. Real-time outputs perceive, recognize and analyze both stati
   images and time-varying patterns for specific objects, their heading
   speed, shading and color differences
  
   By mimicking the eye plus the visual regions of the brain, the GVP
   culls the essential features. So instead of capturing frames o
   pixels, the chip identifies objects of interest, determines eac
   object's speed and direction, then follows them by tracking thei
   color through the scene
  
   The chip emulates the eye, which has 5 million cones sensitive t
   color, only 15 percent of which see "blue" (the rest are red an
   green) and 140 million monochromatic rods that are 35 times mor
   sensitive than the cones
  
   The chip mimics the human eye's two processing steps, tonic an
   phasic. Tonic processing auto-scales according to ambient ligh
   conditions, enabling it to adapt to a range of luminosity. Phasi
   processing determines movement by using local variables in feedbac
   loops. As light's edges pass over the cones and rods of the eye, thes
   local feedback loops detect contrast changes caused by objects movin
   through the scene
  
   For detecting smooth contours, rather than sharp contrast changes, th
   eye adds ocular movement. The eye typically sweeps a scene about tw
   to three times a second as well as making vibratory movements at abou
   100 Hz. The faster jitter accounts for the visual acuity of th
   eye--sensitivity to the smallest detectable feature, which is an edg
   moving between two adjacent rods or cones
  
   After all this processing, the visual signal is then sent to the brai
   for higher-level observation and recognition tasks. Because onl
   detected movement and color along with the shape and contour o
   objects is sent up to the brain, rather than raw pixels, the averag
   compression ratio of information, according to the company, is abou
   145. For the future, Pirim is working on a "visual" mouse for
   hand-gesture interface to computers that takes advantage of that hig
   compression ratio
  
   The electronic eye mimics the theoretical processing steps of the rea
   eye with hard-wired silicon circuitry around each pixel in its senso
   array. Each pixel is analyzed by the vision chip with hardware tha
   determines and scales luminescence, tracks color, remembers movemen
   in the previous moment, remembers the direction of previous movement
   and deduces the speed of the detected objects from parallel phasic an
   tonic neural circuitry
  
   Basically, each parameter has an associated neuron that handles it
   processing tasks in parallel. In addition, each pixel has tw
   auxiliary neurons that define the zone in which an object i
   located--that is, from the direction in which an object is moving
   these neurons deduce the leading and trailing edge of the object an
   mark with registers associated with the first (leading-edge) and las
   (trailing-edge) pixel belonging to the object
  
   Each of these silicon neurons is built with RAM, a few registers, a
   adder and a comparator. The auxiliary neurons also have a multiplie
   at their disposal
  
   Supplied as a 100-pin module, the chip accommodates analog-input lin
   levels for video input, with an input amplifier with programmable gai
   autoscaling the signal
  
   The modules measure 40 mm2, have 100 pins and can handle 20-MHz vide
   signals. The chip is priced at $960. On a card with a socketed GVP
   and 64 kbytes of Flash RAM the price comes to $1,500. A developer'
   daughtercard version, with a motherboard accommodating inputs fo
   video, a power supply and computer interface along with a PAL encode
   and a library of software to manage the unit is available for $4,650
  

-- Lloyd Rasmussen
Senior Staff Engineer, Engineering Section
National Library Service for the Blind and Physically Handicapped
Library of Congress 202-707-0535
(work) lras@loc.gov www.loc.gov/nls/
(home) lras@sprynet.com



This archive was generated by hypermail 2b29 : Sun Dec 02 2012 - 01:30:04 PST